That's the "Facial recognition fooled by funny-colored glasses" study from Carngie Mellon, where researchers were able to make machine learning algorithms fail disastrously (e.g. mistake a man for Milla Jovovich) with a pair of glasses printed with what looks like a random assortment of colorful pixels, but is in fact a targeted attack specifically designed to trick the algorithm.
This Alexa failure is obviously not a targeted attack; but when your system is exposed to enough data, eventually you will stumble over some input that happens to resemble a targeted attack by pure chance, right? It's equivalent to saying that if you aimed the facial recognition algos in that paper at millions of faces wearing randomly-colored glasses, eventually some of the glasses would be close enough to the targeted-attack glasses to produce the same effect.
Obviously I'm theorizing on almost no data here - I don't know anything about Alexa's voice recognition, maybe it contains no ML at all. But it seems plausible that this might be what happened here - not a bug per se, but the natural and totally expected result of giving an opaque, machine-generated system with a very low failure rate so much input data that the failure rate is significant.